Speaking of SSD IOPs. Running the same tests on my SSDs (LiteOn ECT-480N9S 480GB SSDs): The lines at the bottom are a single 6TB spinning disk for comparison's sake. http://imgur.com/a/fD0Mh Based on these numbers, there is a minimum latency per operation, but multiple operations can be performed simultaneously. The sweet spot for my SSDs is ~8 journals per SSD to maximize IOPs on a per journal basis. Unfortunately, at 8 journals, the overall IOPs is much less than the stated IOPs for the SSD. (~5000 vs 9000 IOPs). Better than spinning disks, but not what I was expecting. The spreadsheet is available here: https://people.beocat.cis.ksu.edu/~mozes/hobbit-ssd-vs-std-iops.ods -- Adam On Tue, Mar 31, 2015 at 7:09 AM, fred@xxxxxxxxxx <fred@xxxxxxxxxx> wrote: > Hi, > > in our quest to get the right SSD for OSD journals, I managed to benchmark > two kind of "10 DWPD" SSDs : > - Toshiba M2 PX02SMF020 > - Samsung 845DC PRO > > I wan't to determine if a disk is appropriate considering its absolute > performances, and the optimal number of ceph-osd processes using the SSD as > a journal. > The benchmark consists of a fio command, with SYNC and DIRECT access > options, and 4k blocks write accesses. > > fio --filename=/dev/sda --direct=1 --sync=1 --rw=write --bs=4k --runtime=60 > --time_based --group_reporting --name=journal-test --iodepth=<1 or 16> > --numjobs=< ranging from 1 to 16> > > I think numjobs can represent the concurrent number of OSD served by this > SSD. Am I right on this ? > > > http://www.4shared.com/download/WOvooKVXce/Fio-Direct-Sync-ToshibaM2-Sams.png?lgfp=3000 > > My understanding of that data is that the 845DC Pro cannot be used for more > that 4 OSD. > The M2 is very constant in its comportment. > The iodepth has almost no impact on perfs here. > > Could someone having other SSD types make the same test to consolidate the > data ? > > Among the short list that could be considered for that task (for their > price/perfs/DWPD/...) : > - Seagate 1200 SSD 200GB, SAS 12Gb/s ST200FM0053 > - Hitachi SSD800MM MLC HUSMM8020ASS200 > - Intel DC3700 > > I've not yet considered write amplification mentionned in other posts. > > Frederic > > Josef Johansson <josef86@xxxxxxxxx> a écrit le 20/03/15 10:29 : > > > The 845DC Pro does look really nice, comparable with s3700 with TDW even. > The price is what really does it, as it’s almost a third compared with > s3700.. > > > > > > _______________________________________________ > ceph-users mailing list > ceph-users@xxxxxxxxxxxxxx > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |